[Intel-gfx] [PATCH 09/15] drm/i915: Eliminate lots of iterations over the execobjects array
Chris Wilson
chris at chris-wilson.co.uk
Thu Mar 16 13:20:00 UTC 2017
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
drivers/gpu/drm/i915/i915_drv.h | 2 +-
drivers/gpu/drm/i915/i915_gem_evict.c | 95 +-
drivers/gpu/drm/i915/i915_gem_execbuffer.c | 1683 +++++++++++++----------
drivers/gpu/drm/i915/i915_vma.c | 2 +-
drivers/gpu/drm/i915/i915_vma.h | 1 +
drivers/gpu/drm/i915/selftests/i915_gem_evict.c | 4 +-
6 files changed, 1019 insertions(+), 768 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index bae251806fb6..4bb397d84923 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3534,7 +3534,7 @@ int __must_check i915_gem_evict_something(struct i915_address_space *vm,
int __must_check i915_gem_evict_for_node(struct i915_address_space *vm,
struct drm_mm_node *node,
unsigned int flags);
-int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle);
+int i915_gem_evict_vm(struct i915_address_space *vm);
/* belongs in i915_gem_gtt.h */
static inline void i915_gem_chipset_flush(struct drm_i915_private *dev_priv)
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
index fd5c75517143..c45d58e8a53d 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/i915_gem_evict.c
@@ -50,6 +50,30 @@ static bool ggtt_is_idle(struct drm_i915_private *dev_priv)
return true;
}
+static int ggtt_flush(struct drm_i915_private *i915)
+{
+ int err;
+
+ /* Not everything in the GGTT is tracked via vma (otherwise we
+ * could evict as required with minimal stalling) so we are forced
+ * to idle the GPU and explicitly retire outstanding requests in
+ * the hopes that we can then remove contexts and the like only
+ * bound by their active reference.
+ */
+ err = i915_gem_switch_to_kernel_context(i915);
+ if (err)
+ return err;
+
+ err = i915_gem_wait_for_idle(i915,
+ I915_WAIT_INTERRUPTIBLE |
+ I915_WAIT_LOCKED);
+ if (err)
+ return err;
+
+ i915_gem_retire_requests(i915);
+ return 0;
+}
+
static bool
mark_free(struct drm_mm_scan *scan,
struct i915_vma *vma,
@@ -175,23 +199,10 @@ i915_gem_evict_something(struct i915_address_space *vm,
return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC;
}
- /* Not everything in the GGTT is tracked via vma (otherwise we
- * could evict as required with minimal stalling) so we are forced
- * to idle the GPU and explicitly retire outstanding requests in
- * the hopes that we can then remove contexts and the like only
- * bound by their active reference.
- */
- ret = i915_gem_switch_to_kernel_context(dev_priv);
- if (ret)
- return ret;
-
- ret = i915_gem_wait_for_idle(dev_priv,
- I915_WAIT_INTERRUPTIBLE |
- I915_WAIT_LOCKED);
+ ret = ggtt_flush(dev_priv);
if (ret)
return ret;
- i915_gem_retire_requests(dev_priv);
goto search_again;
found:
@@ -338,10 +349,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
/**
* i915_gem_evict_vm - Evict all idle vmas from a vm
* @vm: Address space to cleanse
- * @do_idle: Boolean directing whether to idle first.
*
- * This function evicts all idles vmas from a vm. If all unpinned vmas should be
- * evicted the @do_idle needs to be set to true.
+ * This function evicts all vmas from a vm.
*
* This is used by the execbuf code as a last-ditch effort to defragment the
* address space.
@@ -349,38 +358,50 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
* To clarify: This is for freeing up virtual address space, not for freeing
* memory in e.g. the shrinker.
*/
-int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle)
+int i915_gem_evict_vm(struct i915_address_space *vm)
{
+ struct list_head *phases[] = {
+ &vm->inactive_list,
+ &vm->active_list,
+ NULL
+ }, **phase;
+ struct list_head eviction_list;
struct i915_vma *vma, *next;
int ret;
lockdep_assert_held(&vm->i915->drm.struct_mutex);
trace_i915_gem_evict_vm(vm);
- if (do_idle) {
- struct drm_i915_private *dev_priv = vm->i915;
-
- if (i915_is_ggtt(vm)) {
- ret = i915_gem_switch_to_kernel_context(dev_priv);
- if (ret)
- return ret;
- }
-
- ret = i915_gem_wait_for_idle(dev_priv,
- I915_WAIT_INTERRUPTIBLE |
- I915_WAIT_LOCKED);
+ /* Switch back to the default context in order to unpin
+ * the existing context objects. However, such objects only
+ * pin themselves inside the global GTT and performing the
+ * switch otherwise is ineffective.
+ */
+ if (i915_is_ggtt(vm)) {
+ ret = ggtt_flush(vm->i915);
if (ret)
return ret;
-
- i915_gem_retire_requests(dev_priv);
- WARN_ON(!list_empty(&vm->active_list));
}
- list_for_each_entry_safe(vma, next, &vm->inactive_list, vm_link)
- if (!i915_vma_is_pinned(vma))
- WARN_ON(i915_vma_unbind(vma));
+ INIT_LIST_HEAD(&eviction_list);
+ phase = phases;
+ do {
+ list_for_each_entry(vma, *phase, vm_link) {
+ if (i915_vma_is_pinned(vma))
+ continue;
+
+ __i915_vma_pin(vma);
+ list_add(&vma->evict_link, &eviction_list);
+ }
+ } while (*++phase);
- return 0;
+ ret = 0;
+ list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+ __i915_vma_unpin(vma);
+ if (ret == 0)
+ ret = i915_vma_unbind(vma);
+ }
+ return ret;
}
#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 3cfad68fc452..5ce4f2e47eb9 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -42,17 +42,134 @@
#define DBG_USE_CPU_RELOC 0 /* -1 force GTT relocs; 1 force CPU relocs */
-#define __EXEC_OBJECT_HAS_PIN (1<<31)
-#define __EXEC_OBJECT_HAS_FENCE (1<<30)
-#define __EXEC_OBJECT_NEEDS_MAP (1<<29)
-#define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
+#define __EXEC_OBJECT_HAS_PIN BIT(31)
+#define __EXEC_OBJECT_HAS_FENCE BIT(30)
+#define __EXEC_OBJECT_NEEDS_MAP BIT(29)
+#define __EXEC_OBJECT_NEEDS_BIAS BIT(28)
#define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
+#define __EB_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE)
+
+#define __EXEC_HAS_RELOC BIT(31)
+#define __EXEC_VALIDATED BIT(30)
+#define UPDATE PIN_OFFSET_FIXED
#define BATCH_OFFSET_BIAS (256*1024)
#define __I915_EXEC_ILLEGAL_FLAGS \
(__I915_EXEC_UNKNOWN_FLAGS | I915_EXEC_CONSTANTS_MASK)
+/**
+ * DOC: User command execution
+ *
+ * Userspace submits commands to be executed on the GPU as an instruction
+ * stream within a GEM object we call a batchbuffer. This instructions may
+ * refer to other GEM objects containing auxiliary state such as kernels,
+ * samplers, render targets and even secondary batchbuffers. Userspace does
+ * not know where in the GPU memory these objects reside and so before the
+ * batchbuffer is passed to the GPU for execution, those addresses in the
+ * batchbuffer and auxiliary objects are updated. This is known as relocation,
+ * or patching. To try and avoid having to relocate each object on the next
+ * execution, userspace is told the location of those objects in this pass,
+ * but this remains just a hint as the kernel may choose a new location for
+ * any object in the future.
+ *
+ * Processing an execbuf ioctl is conceptually split up into a few phases.
+ *
+ * 1. Validation - Ensure all the pointers, handles and flags are valid.
+ * 2. Reservation - Assign GPU address space for every object
+ * 3. Relocation - Update any addresses to point to the final locations
+ * 4. Serialisation - Order the request with respect to its dependencies
+ * 5. Construction - Construct a request to execute the batchbuffer
+ * 6. Submission (at some point in the future execution)
+ *
+ * Reserving resources for the execbuf is the most complicated phase. We
+ * neither want to have to migrate the object in the address space, nor do
+ * we want to have to update any relocations pointing to this object. Ideally,
+ * we want to leave the object where it is and for all the existing relocations
+ * to match. If the object is given a new address, or if userspace thinks the
+ * object is elsewhere, we have to parse all the relocation entries and update
+ * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
+ * all the target addresses in all of its objects match the value in the
+ * relocation entries and that they all match the presumed offsets given by the
+ * list of execbuffer objects. Using this knowledge, we know that if we haven't
+ * moved any buffers, all the relocation entries are valid and we can skip
+ * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
+ * hang.)
+ *
+ * The reservation is done is multiple phases. First we try and keep any
+ * object already bound in its current location - so as long as meets the
+ * constraints imposed by the new execbuffer. Any object left unbound after the
+ * first pass is then fitted into any available idle space. If an object does
+ * not fit, all objects are removed from the reservation and the process rerun
+ * after sorting the objects into a priority order (more difficult to fit
+ * objects are tried first). Failing that, the entire VM is cleared and we try
+ * to fit the execbuf once last time before concluding that it simply will not
+ * fit.
+ *
+ * A small complication to all of this is that we allow userspace not only to
+ * specify an alignment and a size for the object in the address space, but
+ * we also allow userspace to specify the exact offset. This objects are
+ * simpler to place (the location is known a priori) all we have to do is make
+ * sure the space is available.
+ *
+ * Once all the objects are in place, patching up the buried pointers to point
+ * to the final locations is a fairly simple job of walking over the relocation
+ * entry arrays, looking up the right address and rewriting the value into
+ * the object. Simple! ... The relocation entries are stored in user memory
+ * and so to access them we have to copy them into a local buffer. That copy
+ * has to avoid taking any pagefaults as they may lead back to a GEM object
+ * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
+ * the relocation into multiple passes. First we try to do everything within an
+ * atomic context (avoid the pagefaults) which requires that we never wait. If
+ * we detect that we may wait, or if we need to fault, then we have to fallback
+ * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
+ * bells yet?) Dropping the mutex means that we lose all the state we have
+ * built up so far for the execbuf and we must reset any global data. However,
+ * we do leave the objects pinned in their final locations - which is a
+ * potential issue for concurrent execbufs. Once we have left the mutex, we can
+ * allocate and copy all the relocation entries into a large array at our
+ * leisure, reacquire the mutex, reclaim all the objects and other state and
+ * then proceed to update any incorrect addresses with the objects.
+ *
+ * As we process the relocation entries, we maintain a record of whether the
+ * object is being written to. Using NORELOC, we expect userspace to provide
+ * this information instead. We also check whether we can skip the relocation
+ * by comparing the expected value inside the relocation entry with the target's
+ * final address. If they differ, we have to map the current object and rewrite
+ * the 4 or 8 byte pointer within.
+ *
+ * Serialising an execbuf is quite simple according to the rules of the GEM
+ * ABI. Execution within each context is ordered by the order of submission.
+ * Writes to any GEM object are in order of submission and are exclusive. Reads
+ * from a GEM object are unordered with respect to other reads, but ordered by
+ * writes. A write submitted after a read cannot occur before the read, and
+ * similarly any read submitted after a write cannot occur before the write.
+ * Writes are ordered between engines such that only one write occurs at any
+ * time (completing any reads beforehand) - using semaphores where available
+ * and CPU serialisation otherwise. Other GEM access obey the same rules, any
+ * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
+ * reads before starting, and any read (either using set-domain or pread) must
+ * flush all GPU writes before starting. (Note we only employ a barrier before,
+ * we currently rely on userspace not concurrently starting a new execution
+ * whilst reading or writing to an object. This may be an advantage or not
+ * depending on how much you trust userspace not to shoot themselves in the
+ * foot.) Serialisation may just result in the request being inserted into
+ * a DAG awaiting its turn, but most simple is to wait on the CPU until
+ * all dependencies are resolved.
+ *
+ * After all of that, is just a matter of closing the request and handing it to
+ * the hardware (well, leaving it in a queue to be executed). However, we also
+ * offer the ability for batchbuffers to be run with elevated privileges so
+ * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
+ * Before any batch is given extra privileges we first must check that it
+ * contains no nefarious instructions, we check that each instruction is from
+ * our whitelist and all registers are also from an allowed list. We first
+ * copy the user's batchbuffer to a shadow (so that the user doesn't have
+ * access to it, either by the CPU or GPU as we scan it) and then parse each
+ * instruction. If everything is ok, we set a flag telling the hardware to run
+ * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
+ */
+
struct i915_execbuffer {
struct drm_i915_private *i915;
struct drm_file *file;
@@ -63,28 +180,62 @@ struct i915_execbuffer {
struct i915_address_space *vm;
struct i915_vma *batch;
struct drm_i915_gem_request *request;
- u32 batch_start_offset;
- u32 batch_len;
- unsigned int dispatch_flags;
- struct drm_i915_gem_exec_object2 shadow_exec_entry;
- bool need_relocs;
- struct list_head vmas;
+ unsigned int buffer_count;
+ struct list_head unbound;
+ struct list_head relocs;
struct reloc_cache {
struct drm_mm_node node;
unsigned long vaddr;
unsigned int page;
bool use_64bit_reloc;
+ bool has_llc;
+ bool has_fence;
} reloc_cache;
- int lut_mask;
+ u64 invalid_flags;
+ u32 context_flags;
+ u32 dispatch_flags;
+ u32 batch_start_offset;
+ u32 batch_len;
+ int lut_size;
struct hlist_head *buckets;
};
+#define to_ptr(T, x) ((T *)(uintptr_t)(x))
+
+/* Used to convert any address to canonical form.
+ * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
+ * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
+ * addresses to be in a canonical form:
+ * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
+ * canonical form [63:48] == [47]."
+ */
+#define GEN8_HIGH_ADDRESS_BIT 47
+static inline u64 gen8_canonical_addr(u64 address)
+{
+ return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
+}
+
+static inline u64 gen8_noncanonical_addr(u64 address)
+{
+ return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
+}
+
static int
eb_create(struct i915_execbuffer *eb)
{
if ((eb->args->flags & I915_EXEC_HANDLE_LUT) == 0) {
- unsigned int size = 1 + ilog2(eb->args->buffer_count);
-
+ unsigned int size = 1 + ilog2(eb->buffer_count);
+
+ /* Without a 1:1 association between relocation handles and
+ * the execobject[] index, we instead create a hashtable.
+ * We size it dynamically based on available memory, starting
+ * first with 1:1 assocative hash and scaling back until
+ * the allocation succeeds.
+ *
+ * Later on we use a positive lut_size to indicate we are
+ * using this hashtable, and a negative value to indicate a
+ * direct lookup.
+ */
do {
eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
@@ -99,90 +250,349 @@ eb_create(struct i915_execbuffer *eb)
return -ENOMEM;
}
- eb->lut_mask = size;
+ eb->lut_size = size;
} else {
- eb->lut_mask = -eb->args->buffer_count;
+ eb->lut_size = -eb->buffer_count;
}
return 0;
}
+static bool
+eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
+ const struct i915_vma *vma)
+{
+ if ((entry->flags & __EXEC_OBJECT_HAS_PIN) == 0)
+ return true;
+
+ if (vma->node.size < entry->pad_to_size)
+ return true;
+
+ if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
+ return true;
+
+ if (entry->flags & EXEC_OBJECT_PINNED &&
+ vma->node.start != entry->offset)
+ return true;
+
+ if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
+ vma->node.start < BATCH_OFFSET_BIAS)
+ return true;
+
+ if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
+ (vma->node.start + vma->node.size - 1) >> 32)
+ return true;
+
+ return false;
+}
+
+static void
+eb_pin_vma(struct i915_execbuffer *eb,
+ struct drm_i915_gem_exec_object2 *entry,
+ struct i915_vma *vma)
+{
+ u64 flags;
+
+ flags = vma->node.start;
+ flags |= PIN_USER | PIN_NONBLOCK | PIN_OFFSET_FIXED;
+ if (unlikely(entry->flags & EXEC_OBJECT_NEEDS_GTT))
+ flags |= PIN_GLOBAL;
+ if (unlikely(i915_vma_pin(vma, 0, 0, flags)))
+ return;
+
+ if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
+ if (unlikely(i915_vma_get_fence(vma))) {
+ i915_vma_unpin(vma);
+ return;
+ }
+
+ if (i915_vma_pin_fence(vma))
+ entry->flags |= __EXEC_OBJECT_HAS_FENCE;
+ }
+
+ entry->flags |= __EXEC_OBJECT_HAS_PIN;
+}
+
static inline void
__eb_unreserve_vma(struct i915_vma *vma,
const struct drm_i915_gem_exec_object2 *entry)
{
+ GEM_BUG_ON((entry->flags & __EXEC_OBJECT_HAS_PIN) == 0);
+
if (unlikely(entry->flags & __EXEC_OBJECT_HAS_FENCE))
i915_vma_unpin_fence(vma);
- if (entry->flags & __EXEC_OBJECT_HAS_PIN)
- __i915_vma_unpin(vma);
+ __i915_vma_unpin(vma);
}
-static void
-eb_unreserve_vma(struct i915_vma *vma)
+static inline void
+eb_unreserve_vma(struct i915_vma *vma,
+ struct drm_i915_gem_exec_object2 *entry)
{
- struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
-
- __eb_unreserve_vma(vma, entry);
- entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
+ if (entry->flags & __EXEC_OBJECT_HAS_PIN) {
+ __eb_unreserve_vma(vma, entry);
+ entry->flags &= ~__EB_RESERVED;
+ }
}
-static void
-eb_reset(struct i915_execbuffer *eb)
+static int
+eb_add_vma(struct i915_execbuffer *eb,
+ struct drm_i915_gem_exec_object2 *entry,
+ struct i915_vma *vma)
{
- struct i915_vma *vma;
+ int ret;
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- eb_unreserve_vma(vma);
- i915_vma_put(vma);
- vma->exec_entry = NULL;
+ GEM_BUG_ON(i915_vma_is_closed(vma));
+
+ if ((eb->args->flags & __EXEC_VALIDATED) == 0) {
+ if (unlikely(entry->flags & eb->invalid_flags))
+ return -EINVAL;
+
+ if (unlikely(entry->alignment && !is_power_of_2(entry->alignment)))
+ return -EINVAL;
+
+ /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
+ * any non-page-aligned or non-canonical addresses.
+ */
+ if (entry->flags & EXEC_OBJECT_PINNED) {
+ if (unlikely(entry->offset !=
+ gen8_canonical_addr(entry->offset & PAGE_MASK)))
+ return -EINVAL;
+ }
+
+ /* From drm_mm perspective address space is continuous,
+ * so from this point we're always using non-canonical
+ * form internally.
+ */
+ entry->offset = gen8_noncanonical_addr(entry->offset);
+
+ /* pad_to_size was once a reserved field, so sanitize it */
+ if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
+ if (unlikely(offset_in_page(entry->pad_to_size)))
+ return -EINVAL;
+ } else {
+ entry->pad_to_size = 0;
+ }
+
+ if (unlikely(vma->exec_entry)) {
+ DRM_DEBUG("Object [handle %d, index %d] appears more than once in object list\n",
+ entry->handle, (int)(entry - eb->exec));
+ return -EINVAL;
+ }
}
- if (eb->lut_mask >= 0)
- memset(eb->buckets, 0,
- (1<<eb->lut_mask)*sizeof(struct hlist_head));
-}
+ vma->exec_entry = entry;
+ entry->rsvd2 = (uintptr_t)vma;
+ i915_vma_get(vma);
-#define to_ptr(T, x) ((T *)(uintptr_t)(x))
+ if (eb->lut_size >= 0) {
+ vma->exec_handle = entry->handle;
+ hlist_add_head(&vma->exec_node,
+ &eb->buckets[hash_32(entry->handle,
+ eb->lut_size)]);
+ }
-static bool
-eb_add_vma(struct i915_execbuffer *eb, struct i915_vma *vma, int i)
+ if (entry->relocation_count)
+ list_add_tail(&vma->reloc_link, &eb->relocs);
+
+ if (!eb->reloc_cache.has_fence) {
+ entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
+ } else {
+ if (entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
+ i915_gem_object_is_tiled(vma->obj))
+ entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
+ }
+
+ if ((entry->flags & EXEC_OBJECT_PINNED) == 0)
+ entry->flags |= eb->context_flags;
+
+ ret = 0;
+ if (vma->node.size)
+ eb_pin_vma(eb, entry, vma);
+ if (eb_vma_misplaced(entry, vma)) {
+ eb_unreserve_vma(vma, entry);
+
+ list_add_tail(&vma->exec_link, &eb->unbound);
+ if (drm_mm_node_allocated(&vma->node))
+ ret = i915_vma_unbind(vma);
+ } else {
+ if (entry->offset != vma->node.start) {
+ entry->offset = vma->node.start | UPDATE;
+ eb->args->flags |= __EXEC_HAS_RELOC;
+ }
+ }
+ return ret;
+}
+
+static inline int use_cpu_reloc(const struct reloc_cache *cache,
+ const struct drm_i915_gem_object *obj)
{
- if (unlikely(vma->exec_entry)) {
- DRM_DEBUG("Object [handle %d, index %d] appears more than once in object list\n",
- eb->exec[i].handle, i);
+ if (!i915_gem_object_has_struct_page(obj))
return false;
+
+ if (DBG_USE_CPU_RELOC)
+ return DBG_USE_CPU_RELOC > 0;
+
+ return (cache->has_llc ||
+ obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
+ obj->cache_level != I915_CACHE_NONE);
+}
+
+static int
+eb_reserve_vma(struct i915_execbuffer *eb, struct i915_vma *vma)
+{
+ struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
+ u64 flags;
+ int ret;
+
+ flags = PIN_USER | PIN_NONBLOCK;
+ if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
+ flags |= PIN_GLOBAL;
+
+ if (!drm_mm_node_allocated(&vma->node)) {
+ /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
+ * limit address to the first 4GBs for unflagged objects.
+ */
+ if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
+ flags |= PIN_ZONE_4G;
+
+ if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
+ flags |= PIN_MAPPABLE;
+
+ if (entry->flags & EXEC_OBJECT_PINNED) {
+ flags |= entry->offset | PIN_OFFSET_FIXED;
+ /* force overlapping PINNED checks */
+ flags &= ~PIN_NONBLOCK;
+ } else if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
+ flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
}
- list_add_tail(&vma->exec_link, &eb->vmas);
- vma->exec_entry = &eb->exec[i];
- if (eb->lut_mask >= 0) {
- vma->exec_handle = eb->exec[i].handle;
- hlist_add_head(&vma->exec_node,
- &eb->buckets[hash_32(vma->exec_handle,
- eb->lut_mask)]);
+ ret = i915_vma_pin(vma, entry->pad_to_size, entry->alignment, flags);
+ if (ret)
+ return ret;
+
+ if (entry->offset != vma->node.start) {
+ entry->offset = vma->node.start | UPDATE;
+ eb->args->flags |= __EXEC_HAS_RELOC;
}
+ entry->flags |= __EXEC_OBJECT_HAS_PIN;
- i915_vma_get(vma);
- eb->exec[i].rsvd2 = (uintptr_t)vma;
- return true;
+ if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
+ ret = i915_vma_get_fence(vma);
+ if (ret)
+ return ret;
+
+ if (i915_vma_pin_fence(vma))
+ entry->flags |= __EXEC_OBJECT_HAS_FENCE;
+ }
+
+ GEM_BUG_ON(eb_vma_misplaced(entry, vma));
+ return 0;
}
-static inline struct hlist_head *ht_head(struct i915_gem_context *ctx,
- u32 handle)
+static int eb_reserve(struct i915_execbuffer *eb)
+{
+ const unsigned int count = eb->buffer_count;
+ struct list_head last;
+ struct i915_vma *vma;
+ unsigned int i, pass;
+ int ret;
+
+ /* Attempt to pin all of the buffers into the GTT.
+ * This is done in 3 phases:
+ *
+ * 1a. Unbind all objects that do not match the GTT constraints for
+ * the execbuffer (fenceable, mappable, alignment etc).
+ * 1b. Increment pin count for already bound objects.
+ * 2. Bind new objects.
+ * 3. Decrement pin count.
+ *
+ * This avoid unnecessary unbinding of later objects in order to make
+ * room for the earlier objects *unless* we need to defragment.
+ */
+
+ pass = 0;
+ ret = 0;
+ do {
+ list_for_each_entry(vma, &eb->unbound, exec_link) {
+ ret = eb_reserve_vma(eb, vma);
+ if (ret)
+ break;
+ }
+ if (ret != -ENOSPC || pass++)
+ return ret;
+
+ /* Resort *all* the objects into priority order */
+ INIT_LIST_HEAD(&eb->unbound);
+ INIT_LIST_HEAD(&last);
+ for (i = 0; i < count; i++) {
+ struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+
+ vma = to_ptr(struct i915_vma, entry->rsvd2);
+ eb_unreserve_vma(vma, entry);
+
+ if (entry->flags & EXEC_OBJECT_PINNED)
+ list_add(&vma->exec_link, &eb->unbound);
+ else if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
+ list_add_tail(&vma->exec_link, &eb->unbound);
+ else
+ list_add_tail(&vma->exec_link, &last);
+ }
+ list_splice_tail(&last, &eb->unbound);
+
+ /* Too fragmented, unbind everything and retry */
+ ret = i915_gem_evict_vm(eb->vm);
+ if (ret)
+ return ret;
+ } while (1);
+}
+
+static inline struct hlist_head *
+ht_head(const struct i915_gem_context *ctx, u32 handle)
{
return &ctx->vma.ht[hash_32(handle, ctx->vma.ht_bits)];
}
+static int eb_batch_index(const struct i915_execbuffer *eb)
+{
+ return eb->buffer_count - 1;
+}
+
+static int eb_select_context(struct i915_execbuffer *eb)
+{
+ struct i915_gem_context *ctx;
+
+ ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
+ if (unlikely(IS_ERR(ctx)))
+ return PTR_ERR(ctx);
+
+ if (unlikely(i915_gem_context_is_banned(ctx))) {
+ DRM_DEBUG("Context %u tried to submit while banned\n",
+ ctx->user_handle);
+ return -EIO;
+ }
+
+ eb->ctx = i915_gem_context_get(ctx);
+ eb->vm = ctx->ppgtt ? &ctx->ppgtt->base : &eb->i915->ggtt.base;
+
+ eb->context_flags = 0;
+ if (ctx->flags & CONTEXT_NO_ZEROMAP)
+ eb->context_flags |= __EXEC_OBJECT_NEEDS_BIAS;
+
+ return 0;
+}
+
static int
eb_lookup_vmas(struct i915_execbuffer *eb)
{
- const int count = eb->args->buffer_count;
+ const int count = eb->buffer_count;
struct i915_vma *vma;
+ struct idr *idr;
int slow_pass = -1;
- int i;
+ int i, ret;
- INIT_LIST_HEAD(&eb->vmas);
+ INIT_LIST_HEAD(&eb->relocs);
+ INIT_LIST_HEAD(&eb->unbound);
if (unlikely(eb->ctx->vma.ht_size & 1))
flush_work(&eb->ctx->vma.resize);
@@ -195,8 +605,9 @@ eb_lookup_vmas(struct i915_execbuffer *eb)
if (vma->ctx_handle != eb->exec[i].handle)
continue;
- if (!eb_add_vma(eb, vma, i))
- return -EINVAL;
+ ret = eb_add_vma(eb, &eb->exec[i], vma);
+ if (unlikely(ret))
+ return ret;
goto next_vma;
}
@@ -207,24 +618,25 @@ next_vma: ;
}
if (slow_pass < 0)
- return 0;
+ goto out;
spin_lock(&eb->file->table_lock);
/* Grab a reference to the object and release the lock so we can lookup
* or create the VMA without using GFP_ATOMIC */
+ idr = &eb->file->object_idr;
for (i = slow_pass; i < count; i++) {
struct drm_i915_gem_object *obj;
if (eb->exec[i].rsvd2)
continue;
- obj = to_intel_bo(idr_find(&eb->file->object_idr,
- eb->exec[i].handle));
+ obj = to_intel_bo(idr_find(idr, eb->exec[i].handle));
if (unlikely(!obj)) {
spin_unlock(&eb->file->table_lock);
DRM_DEBUG("Invalid object handle %d at index %d\n",
eb->exec[i].handle, i);
- return -ENOENT;
+ ret = -ENOENT;
+ goto err;
}
eb->exec[i].rsvd2 = 1 | (uintptr_t)obj;
@@ -245,11 +657,12 @@ next_vma: ;
* from the (obj, vm) we don't run the risk of creating
* duplicated vmas for the same vm.
*/
- obj = to_ptr(struct drm_i915_gem_object, eb->exec[i].rsvd2 & ~1);
+ obj = to_ptr(typeof(*obj), eb->exec[i].rsvd2 & ~1);
vma = i915_vma_instance(obj, eb->vm, NULL);
if (unlikely(IS_ERR(vma))) {
DRM_DEBUG("Failed to lookup VMA\n");
- return PTR_ERR(vma);
+ ret = PTR_ERR(vma);
+ goto err;
}
/* First come, first served */
@@ -265,8 +678,9 @@ next_vma: ;
}
}
- if (!eb_add_vma(eb, vma, i))
- return -EINVAL;
+ ret = eb_add_vma(eb, &eb->exec[i], vma);
+ if (unlikely(ret))
+ goto err;
}
if (4*eb->ctx->vma.ht_count > 3*eb->ctx->vma.ht_size ||
4*eb->ctx->vma.ht_count < eb->ctx->vma.ht_size) {
@@ -274,15 +688,10 @@ next_vma: ;
queue_work(system_highpri_wq, &eb->ctx->vma.resize);
}
- return 0;
-}
-
-static struct i915_vma *
-eb_get_batch(struct i915_execbuffer *eb)
-{
- struct i915_vma *vma;
-
- vma = to_ptr(struct i915_vma, eb->exec[eb->args->buffer_count-1].rsvd2);
+out:
+ /* take note of the batch buffer before we might reorder the lists */
+ i = eb_batch_index(eb);
+ eb->batch = to_ptr(struct i915_vma, eb->exec[i].rsvd2);
/*
* SNA is doing fancy tricks with compressing batch buffers, which leads
@@ -293,24 +702,34 @@ eb_get_batch(struct i915_execbuffer *eb)
* Note that actual hangs have only been observed on gen7, but for
* paranoia do it everywhere.
*/
- if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
- vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
+ if ((eb->exec[i].flags & EXEC_OBJECT_PINNED) == 0)
+ eb->exec[i].flags |= __EXEC_OBJECT_NEEDS_BIAS;
+ if (eb->reloc_cache.has_fence)
+ eb->exec[i].flags |= EXEC_OBJECT_NEEDS_FENCE;
- return vma;
+ eb->args->flags |= __EXEC_VALIDATED;
+ return eb_reserve(eb);
+
+err:
+ for (i = slow_pass; i < count; i++) {
+ if (eb->exec[i].rsvd2 & 1)
+ eb->exec[i].rsvd2 = 0;
+ }
+ return ret;
}
static struct i915_vma *
-eb_get_vma(struct i915_execbuffer *eb, unsigned long handle)
+eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle)
{
- if (eb->lut_mask < 0) {
- if (handle >= -eb->lut_mask)
+ if (eb->lut_size < 0) {
+ if (handle >= -eb->lut_size)
return NULL;
return to_ptr(struct i915_vma, eb->exec[handle].rsvd2);
} else {
struct hlist_head *head;
struct i915_vma *vma;
- head = &eb->buckets[hash_32(handle, eb->lut_mask)];
+ head = &eb->buckets[hash_32(handle, eb->lut_size)];
hlist_for_each_entry(vma, head, exec_node) {
if (vma->exec_handle == handle)
return vma;
@@ -319,61 +738,60 @@ eb_get_vma(struct i915_execbuffer *eb, unsigned long handle)
}
}
-static void eb_destroy(struct i915_execbuffer *eb)
+static void
+eb_reset(const struct i915_execbuffer *eb)
{
- struct i915_vma *vma;
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- if (!vma->exec_entry)
- continue;
+ for (i = 0; i < count; i++) {
+ struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+ struct i915_vma *vma = to_ptr(struct i915_vma, entry->rsvd2);
- __eb_unreserve_vma(vma, vma->exec_entry);
+ eb_unreserve_vma(vma, entry);
vma->exec_entry = NULL;
i915_vma_put(vma);
}
- i915_gem_context_put(eb->ctx);
-
- if (eb->lut_mask >= 0)
- kfree(eb->buckets);
+ if (eb->lut_size >= 0)
+ memset(eb->buckets, 0,
+ sizeof(struct hlist_head) << eb->lut_size);
}
-static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
+static void eb_release_vma(const struct i915_execbuffer *eb)
{
- if (!i915_gem_object_has_struct_page(obj))
- return false;
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
- if (DBG_USE_CPU_RELOC)
- return DBG_USE_CPU_RELOC > 0;
+ if (!eb->exec)
+ return;
- return (HAS_LLC(to_i915(obj->base.dev)) ||
- obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
- obj->cache_level != I915_CACHE_NONE);
-}
+ for (i = 0; i < count; i++) {
+ struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+ struct i915_vma *vma = to_ptr(struct i915_vma, entry->rsvd2);
-/* Used to convert any address to canonical form.
- * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
- * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
- * addresses to be in a canonical form:
- * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
- * canonical form [63:48] == [47]."
- */
-#define GEN8_HIGH_ADDRESS_BIT 47
-static inline uint64_t gen8_canonical_addr(uint64_t address)
-{
- return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
+ if (!vma || !vma->exec_entry)
+ continue;
+
+ GEM_BUG_ON(vma->exec_entry != entry);
+ if (entry->flags & __EXEC_OBJECT_HAS_PIN)
+ __eb_unreserve_vma(vma, entry);
+ vma->exec_entry = NULL;
+ i915_vma_put(vma);
+ }
}
-static inline uint64_t gen8_noncanonical_addr(uint64_t address)
+static void eb_destroy(const struct i915_execbuffer *eb)
{
- return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
+ if (eb->lut_size >= 0)
+ kfree(eb->buckets);
}
-static inline uint64_t
+static inline u64
relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
- uint64_t target_offset)
+ const struct i915_vma *target)
{
- return gen8_canonical_addr((int)reloc->delta + target_offset);
+ return gen8_canonical_addr((int)reloc->delta + target->node.start);
}
static void reloc_cache_init(struct reloc_cache *cache,
@@ -382,6 +800,8 @@ static void reloc_cache_init(struct reloc_cache *cache,
cache->page = -1;
cache->vaddr = 0;
/* Must be a variable in the struct to allow GCC to unroll. */
+ cache->has_llc = HAS_LLC(i915);
+ cache->has_fence = INTEL_GEN(i915) < 4;
cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
cache->node.allocated = false;
}
@@ -484,7 +904,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
struct i915_vma *vma;
int ret;
- if (use_cpu_reloc(obj))
+ if (use_cpu_reloc(cache, obj))
return NULL;
ret = i915_gem_object_set_to_gtt_domain(obj, true);
@@ -573,25 +993,26 @@ static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
*addr = value;
}
-static int
-relocate_entry(struct drm_i915_gem_object *obj,
+static u64
+relocate_entry(struct i915_vma *vma,
const struct drm_i915_gem_relocation_entry *reloc,
- struct reloc_cache *cache,
- u64 target_offset)
+ struct i915_execbuffer *eb,
+ const struct i915_vma *target)
{
+ struct drm_i915_gem_object *obj = vma->obj;
u64 offset = reloc->offset;
- bool wide = cache->use_64bit_reloc;
+ u64 target_offset = relocation_target(reloc, target);
+ bool wide = eb->reloc_cache.use_64bit_reloc;
void *vaddr;
- target_offset = relocation_target(reloc, target_offset);
repeat:
- vaddr = reloc_vaddr(obj, cache, offset >> PAGE_SHIFT);
+ vaddr = reloc_vaddr(obj, &eb->reloc_cache, offset >> PAGE_SHIFT);
if (IS_ERR(vaddr))
return PTR_ERR(vaddr);
clflush_write32(vaddr + offset_in_page(offset),
lower_32_bits(target_offset),
- cache->vaddr);
+ eb->reloc_cache.vaddr);
if (wide) {
offset += sizeof(u32);
@@ -600,16 +1021,15 @@ relocate_entry(struct drm_i915_gem_object *obj,
goto repeat;
}
- return 0;
+ return gen8_canonical_addr(target->node.start) | 1;
}
-static int
-eb_relocate_entry(struct i915_vma *vma,
- struct i915_execbuffer *eb,
- struct drm_i915_gem_relocation_entry *reloc)
+static u64
+eb_relocate_entry(struct i915_execbuffer *eb,
+ struct i915_vma *vma,
+ const struct drm_i915_gem_relocation_entry *reloc)
{
struct i915_vma *target;
- u64 target_offset;
int ret;
/* we've already hold a reference to all valid objects */
@@ -640,26 +1060,28 @@ eb_relocate_entry(struct i915_vma *vma,
return -EINVAL;
}
- if (reloc->write_domain)
+ if (reloc->write_domain) {
target->exec_entry->flags |= EXEC_OBJECT_WRITE;
- /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
- * pipe_control writes because the gpu doesn't properly redirect them
- * through the ppgtt for non_secure batchbuffers.
- */
- if (unlikely(IS_GEN6(eb->i915) &&
- reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
- ret = i915_vma_bind(target, target->obj->cache_level,
- PIN_GLOBAL);
- if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
- return ret;
+ /* Sandybridge PPGTT errata: We need a global gtt mapping
+ * for MI and pipe_control writes because the gpu doesn't
+ * properly redirect them through the ppgtt for non_secure
+ * batchbuffers.
+ */
+ if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
+ IS_GEN6(eb->i915)) {
+ ret = i915_vma_bind(target, target->obj->cache_level,
+ PIN_GLOBAL);
+ if (WARN_ONCE(ret,
+ "Unexpected failure to bind target VMA!"))
+ return ret;
+ }
}
/* If the relocation already has the right value in it, no
* more work needs to be done.
*/
- target_offset = gen8_canonical_addr(target->node.start);
- if (target_offset == reloc->presumed_offset)
+ if (gen8_canonical_addr(target->node.start) == reloc->presumed_offset)
return 0;
/* Check that the relocation address is valid... */
@@ -680,34 +1102,33 @@ eb_relocate_entry(struct i915_vma *vma,
return -EINVAL;
}
- ret = relocate_entry(vma->obj, reloc, &eb->reloc_cache, target_offset);
- if (ret)
- return ret;
-
/* and update the user's relocation entry */
- reloc->presumed_offset = target_offset;
- return 0;
+ return relocate_entry(vma, reloc, eb, target);
}
-static int eb_relocate_vma(struct i915_vma *vma, struct i915_execbuffer *eb)
+static int eb_relocate_vma(struct i915_execbuffer *eb, struct i915_vma *vma)
{
#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
- struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
- struct drm_i915_gem_relocation_entry __user *user_relocs;
- struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
- int remain, ret = 0;
-
- user_relocs = u64_to_user_ptr(entry->relocs_ptr);
+ struct drm_i915_gem_relocation_entry stack[N_RELOC(512)];
+ struct drm_i915_gem_relocation_entry __user *urelocs;
+ const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
+ unsigned int remain;
+ urelocs = u64_to_user_ptr(entry->relocs_ptr);
remain = entry->relocation_count;
- while (remain) {
- struct drm_i915_gem_relocation_entry *r = stack_reloc;
- unsigned long unwritten;
- unsigned int count;
+ if (unlikely(remain > ULONG_MAX / sizeof(*urelocs)))
+ return -EINVAL;
- count = min_t(unsigned int, remain, ARRAY_SIZE(stack_reloc));
- remain -= count;
+ /*
+ * We must check that the entire relocation array is safe
+ * to read. However, if the array is not writable the user loses
+ * the updated relocation values.
+ */
+ do {
+ struct drm_i915_gem_relocation_entry *r = stack;
+ unsigned int count =
+ min_t(unsigned int, remain, ARRAY_SIZE(stack));
/* This is the fast path and we cannot handle a pagefault
* whilst holding the struct mutex lest the user pass in the
* relocations contained within a mmaped bo. For in such a case
@@ -716,66 +1137,66 @@ static int eb_relocate_vma(struct i915_vma *vma, struct i915_execbuffer *eb)
* this is bad and so lockdep complains vehemently.
*/
pagefault_disable();
- unwritten = __copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]));
- pagefault_enable();
- if (unlikely(unwritten)) {
- ret = -EFAULT;
+ if (__copy_from_user_inatomic(r, urelocs, count*sizeof(r[0]))) {
+ pagefault_enable();
+ remain = -EFAULT;
goto out;
}
+ pagefault_enable();
+ remain -= count;
do {
- u64 offset = r->presumed_offset;
+ u64 offset = eb_relocate_entry(eb, vma, r);
- ret = eb_relocate_entry(vma, eb, r);
- if (ret)
+ if (likely(offset == 0)) {
+ } else if ((s64)offset < 0) {
+ remain = (s64)offset;
goto out;
-
- if (r->presumed_offset != offset) {
+ } else {
+ /* Note that reporting an error now
+ * leaves everything in an inconsistent
+ * state as we have *already* changed
+ * the relocation value inside the
+ * object. As we have not changed the
+ * reloc.presumed_offset or will not
+ * change the execobject.offset, on the
+ * call we may not rewrite the value
+ * inside the object, leaving it
+ * dangling and causing a GPU hang.
+ */
pagefault_disable();
- unwritten = __put_user(r->presumed_offset,
- &user_relocs->presumed_offset);
+ __put_user(offset & ~1,
+ &urelocs[r-stack].presumed_offset);
pagefault_enable();
- if (unlikely(unwritten)) {
- /* Note that reporting an error now
- * leaves everything in an inconsistent
- * state as we have *already* changed
- * the relocation value inside the
- * object. As we have not changed the
- * reloc.presumed_offset or will not
- * change the execobject.offset, on the
- * call we may not rewrite the value
- * inside the object, leaving it
- * dangling and causing a GPU hang.
- */
- ret = -EFAULT;
- goto out;
- }
}
-
- user_relocs++;
- r++;
- } while (--count);
- }
-
+ } while (r++, --count);
+ urelocs += ARRAY_SIZE(stack);
+ } while (remain);
out:
reloc_cache_reset(&eb->reloc_cache);
- return ret;
+ return remain;
#undef N_RELOC
}
static int
-eb_relocate_vma_slow(struct i915_vma *vma,
- struct i915_execbuffer *eb,
- struct drm_i915_gem_relocation_entry *relocs)
+eb_relocate_vma_slow(struct i915_execbuffer *eb, struct i915_vma *vma)
{
const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
- int i, ret = 0;
+ struct drm_i915_gem_relocation_entry *relocs =
+ to_ptr(typeof(*relocs), entry->relocs_ptr);
+ unsigned int i;
+ int ret;
for (i = 0; i < entry->relocation_count; i++) {
- ret = eb_relocate_entry(vma, eb, &relocs[i]);
- if (ret)
- break;
+ u64 offset = eb_relocate_entry(eb, vma, &relocs[i]);
+
+ if ((s64)offset < 0) {
+ ret = (s64)offset;
+ goto err;
+ }
}
+ ret = 0;
+err:
reloc_cache_reset(&eb->reloc_cache);
return ret;
}
@@ -783,299 +1204,188 @@ eb_relocate_vma_slow(struct i915_vma *vma,
static int eb_relocate(struct i915_execbuffer *eb)
{
struct i915_vma *vma;
- int ret = 0;
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- ret = eb_relocate_vma(vma, eb);
+ /* The objects are in their final locations, apply the relocations. */
+ list_for_each_entry(vma, &eb->relocs, reloc_link) {
+ int ret = eb_relocate_vma(eb, vma);
if (ret)
- break;
+ return ret;
}
- return ret;
-}
-
-static bool only_mappable_for_reloc(unsigned int flags)
-{
- return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
- __EXEC_OBJECT_NEEDS_MAP;
+ return 0;
}
-static int
-eb_reserve_vma(struct i915_vma *vma,
- struct intel_engine_cs *engine,
- bool *need_reloc)
+static int check_relocations(const struct drm_i915_gem_exec_object2 *entry)
{
- struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
- uint64_t flags;
- int ret;
-
- flags = PIN_USER;
- if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
- flags |= PIN_GLOBAL;
+ const unsigned long relocs_max =
+ ULONG_MAX / sizeof(struct drm_i915_gem_relocation_entry);
+ const char __user *addr, *end;
+ unsigned long size;
+ char __maybe_unused c;
+
+ size = entry->relocation_count;
+ if (size == 0)
+ return 0;
- if (!drm_mm_node_allocated(&vma->node)) {
- /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
- * limit address to the first 4GBs for unflagged objects.
- */
- if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
- flags |= PIN_ZONE_4G;
- if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
- flags |= PIN_GLOBAL | PIN_MAPPABLE;
- if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
- flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
- if (entry->flags & EXEC_OBJECT_PINNED)
- flags |= entry->offset | PIN_OFFSET_FIXED;
- if ((flags & PIN_MAPPABLE) == 0)
- flags |= PIN_HIGH;
- }
-
- ret = i915_vma_pin(vma,
- entry->pad_to_size,
- entry->alignment,
- flags);
- if ((ret == -ENOSPC || ret == -E2BIG) &&
- only_mappable_for_reloc(entry->flags))
- ret = i915_vma_pin(vma,
- entry->pad_to_size,
- entry->alignment,
- flags & ~PIN_MAPPABLE);
- if (ret)
- return ret;
+ if (size > relocs_max)
+ return -EINVAL;
- entry->flags |= __EXEC_OBJECT_HAS_PIN;
+ addr = u64_to_user_ptr(entry->relocs_ptr);
+ size *= sizeof(struct drm_i915_gem_relocation_entry);
+ if (!access_ok(VERIFY_WRITE, addr, size))
+ return -EFAULT;
- if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
- ret = i915_vma_get_fence(vma);
+ end = addr + size;
+ for (; addr < end; addr += PAGE_SIZE) {
+ int ret = __get_user(c, addr);
if (ret)
return ret;
-
- if (i915_vma_pin_fence(vma))
- entry->flags |= __EXEC_OBJECT_HAS_FENCE;
}
-
- if (entry->offset != vma->node.start) {
- entry->offset = vma->node.start;
- *need_reloc = true;
- }
-
- return 0;
+ return __get_user(c, end - 1);
}
-static bool
-need_reloc_mappable(struct i915_vma *vma)
+static int
+eb_copy_relocations(const struct i915_execbuffer *eb)
{
- struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
-
- if (entry->relocation_count == 0)
- return false;
-
- if (!i915_vma_is_ggtt(vma))
- return false;
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
+ int ret;
- /* See also use_cpu_reloc() */
- if (HAS_LLC(to_i915(vma->obj->base.dev)))
- return false;
+ for (i = 0; i < count; i++) {
+ struct drm_i915_gem_relocation_entry __user *urelocs;
+ struct drm_i915_gem_relocation_entry *relocs;
+ unsigned int nreloc = eb->exec[i].relocation_count, j;
+ unsigned long size;
- if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
- return false;
+ if (nreloc == 0)
+ continue;
- return true;
-}
+ ret = check_relocations(&eb->exec[i]);
+ if (ret)
+ goto err;
-static bool
-eb_vma_misplaced(struct i915_vma *vma)
-{
- struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
+ urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr);
+ size = nreloc * sizeof(*relocs);
- WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
- !i915_vma_is_ggtt(vma));
+ relocs = drm_malloc_gfp(size, 1, GFP_TEMPORARY);
+ if (!relocs) {
+ ret = -ENOMEM;
+ goto err;
+ }
- if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
- return true;
+ /* copy_from_user is limited to 4GiB */
+ j = 0;
+ do {
+ u32 len = min_t(u64, 1ull<<31, size);
- if (vma->node.size < entry->pad_to_size)
- return true;
+ if (__copy_from_user(relocs + j, urelocs + j, len)) {
+ ret = -EFAULT;
+ goto err;
+ }
- if (entry->flags & EXEC_OBJECT_PINNED &&
- vma->node.start != entry->offset)
- return true;
+ size -= len;
+ BUILD_BUG_ON_NOT_POWER_OF_2(sizeof(*relocs));
+ j += len / sizeof(*relocs);
+ } while (size);
- if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
- vma->node.start < BATCH_OFFSET_BIAS)
- return true;
+ /* As we do not update the known relocation offsets after
+ * relocating (due to the complexities in lock handling),
+ * we need to mark them as invalid now so that we force the
+ * relocation processing next time. Just in case the target
+ * object is evicted and then rebound into its old
+ * presumed_offset before the next execbuffer - if that
+ * happened we would make the mistake of assuming that the
+ * relocations were valid.
+ */
+ user_access_begin();
+ for (j = 0; j < nreloc; j++)
+ unsafe_put_user(-1,
+ &urelocs[j].presumed_offset,
+ end_user);
+end_user:
+ user_access_end();
- /* avoid costly ping-pong once a batch bo ended up non-mappable */
- if (entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
- !i915_vma_is_map_and_fenceable(vma))
- return !only_mappable_for_reloc(entry->flags);
+ eb->exec[i].relocs_ptr = (uintptr_t)relocs;
+ }
- if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
- (vma->node.start + vma->node.size - 1) >> 32)
- return true;
+ return 0;
- return false;
+err:
+ while (i--) {
+ struct drm_i915_gem_relocation_entry *relocs =
+ to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr);
+ if (eb->exec[i].relocation_count)
+ drm_free_large(relocs);
+ }
+ return ret;
}
-static int eb_reserve(struct i915_execbuffer *eb)
+static int eb_prefault_relocations(const struct i915_execbuffer *eb)
{
- const bool has_fenced_gpu_access = INTEL_GEN(eb->i915) < 4;
- struct i915_vma *vma;
- struct list_head ordered_vmas;
- struct list_head pinned_vmas;
- int retry;
-
- INIT_LIST_HEAD(&ordered_vmas);
- INIT_LIST_HEAD(&pinned_vmas);
- while (!list_empty(&eb->vmas)) {
- struct drm_i915_gem_exec_object2 *entry;
- bool need_fence, need_mappable;
-
- vma = list_first_entry(&eb->vmas, struct i915_vma, exec_link);
- entry = vma->exec_entry;
-
- if (eb->ctx->flags & CONTEXT_NO_ZEROMAP)
- entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
-
- if (!has_fenced_gpu_access)
- entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
- need_fence =
- entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
- i915_gem_object_is_tiled(vma->obj);
- need_mappable = need_fence || need_reloc_mappable(vma);
-
- if (entry->flags & EXEC_OBJECT_PINNED)
- list_move_tail(&vma->exec_link, &pinned_vmas);
- else if (need_mappable) {
- entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
- list_move(&vma->exec_link, &ordered_vmas);
- } else
- list_move_tail(&vma->exec_link, &ordered_vmas);
- }
- list_splice(&ordered_vmas, &eb->vmas);
- list_splice(&pinned_vmas, &eb->vmas);
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
- /* Attempt to pin all of the buffers into the GTT.
- * This is done in 3 phases:
- *
- * 1a. Unbind all objects that do not match the GTT constraints for
- * the execbuffer (fenceable, mappable, alignment etc).
- * 1b. Increment pin count for already bound objects.
- * 2. Bind new objects.
- * 3. Decrement pin count.
- *
- * This avoid unnecessary unbinding of later objects in order to make
- * room for the earlier objects *unless* we need to defragment.
- */
- retry = 0;
- do {
- int ret = 0;
-
- /* Unbind any ill-fitting objects or pin. */
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- if (!drm_mm_node_allocated(&vma->node))
- continue;
-
- if (eb_vma_misplaced(vma))
- ret = i915_vma_unbind(vma);
- else
- ret = eb_reserve_vma(vma, eb->engine, &eb->need_relocs);
- if (ret)
- goto err;
- }
-
- /* Bind fresh objects */
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- if (drm_mm_node_allocated(&vma->node))
- continue;
-
- ret = eb_reserve_vma(vma, eb->engine, &eb->need_relocs);
- if (ret)
- goto err;
- }
-
-err:
- if (ret != -ENOSPC || retry++)
- return ret;
+ if (unlikely(i915.prefault_disable))
+ return 0;
- /* Decrement pin count for bound objects */
- list_for_each_entry(vma, &eb->vmas, exec_link)
- eb_unreserve_vma(vma);
+ for (i = 0; i < count; i++) {
+ int ret;
- ret = i915_gem_evict_vm(eb->vm, true);
+ ret = check_relocations(&eb->exec[i]);
if (ret)
return ret;
- } while (1);
+ }
+
+ return 0;
}
-static int
-eb_relocate_slow(struct i915_execbuffer *eb)
+static int eb_relocate_slow(struct i915_execbuffer *eb)
{
- const unsigned int count = eb->args->buffer_count;
struct drm_device *dev = &eb->i915->drm;
- struct drm_i915_gem_relocation_entry *reloc;
+ bool have_copy = false;
struct i915_vma *vma;
- int *reloc_offset;
- int i, total, ret;
+ int ret = 0;
+
+repeat:
+ if (signal_pending(current)) {
+ ret = -ERESTARTSYS;
+ goto out;
+ }
/* We may process another execbuffer during the unlock... */
eb_reset(eb);
mutex_unlock(&dev->struct_mutex);
- total = 0;
- for (i = 0; i < count; i++)
- total += eb->exec[i].relocation_count;
-
- reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
- reloc = drm_malloc_ab(total, sizeof(*reloc));
- if (reloc == NULL || reloc_offset == NULL) {
- drm_free_large(reloc);
- drm_free_large(reloc_offset);
- mutex_lock(&dev->struct_mutex);
- return -ENOMEM;
- }
-
- total = 0;
- for (i = 0; i < count; i++) {
- struct drm_i915_gem_relocation_entry __user *user_relocs;
- u64 invalid_offset = (u64)-1;
- int j;
-
- user_relocs = u64_to_user_ptr(eb->exec[i].relocs_ptr);
-
- if (copy_from_user(reloc+total, user_relocs,
- eb->exec[i].relocation_count * sizeof(*reloc))) {
- ret = -EFAULT;
- mutex_lock(&dev->struct_mutex);
- goto err;
- }
-
- /* As we do not update the known relocation offsets after
- * relocating (due to the complexities in lock handling),
- * we need to mark them as invalid now so that we force the
- * relocation processing next time. Just in case the target
- * object is evicted and then rebound into its old
- * presumed_offset before the next execbuffer - if that
- * happened we would make the mistake of assuming that the
- * relocations were valid.
- */
- for (j = 0; j < eb->exec[i].relocation_count; j++) {
- if (__copy_to_user(&user_relocs[j].presumed_offset,
- &invalid_offset,
- sizeof(invalid_offset))) {
- ret = -EFAULT;
- mutex_lock(&dev->struct_mutex);
- goto err;
- }
- }
-
- reloc_offset[i] = total;
- total += eb->exec[i].relocation_count;
+ /* We take 3 passes through the slowpatch.
+ *
+ * 1 - we try to just prefault all the user relocation entries and
+ * then attempt to reuse the atomic pagefault disabled fast path again.
+ *
+ * 2 - we copy the user entries to a local buffer here outside of the
+ * local and allow ourselves to wait upon any rendering before
+ * relocations
+ *
+ * 3 - we already have a local copy of the relocation entries, but
+ * were interrupted (EAGAIN) whilst waiting for the objects, try again.
+ */
+ if (ret == 0) {
+ ret = eb_prefault_relocations(eb);
+ } else if (!have_copy) {
+ ret = eb_copy_relocations(eb);
+ have_copy = ret == 0;
+ } else {
+ cond_resched();
+ ret = 0;
+ }
+ if (ret) {
+ mutex_lock(&dev->struct_mutex);
+ goto out;
}
ret = i915_mutex_lock_interruptible(dev);
if (ret) {
mutex_lock(&dev->struct_mutex);
- goto err;
+ goto out;
}
/* reacquire the objects */
@@ -1083,16 +1393,18 @@ eb_relocate_slow(struct i915_execbuffer *eb)
if (ret)
goto err;
- ret = eb_reserve(eb);
- if (ret)
- goto err;
-
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- int idx = vma->exec_entry - eb->exec;
-
- ret = eb_relocate_vma_slow(vma, eb, reloc + reloc_offset[idx]);
- if (ret)
- goto err;
+ list_for_each_entry(vma, &eb->relocs, reloc_link) {
+ if (!have_copy) {
+ pagefault_disable();
+ ret = eb_relocate_vma(eb, vma);
+ pagefault_enable();
+ if (ret)
+ goto repeat;
+ } else {
+ ret = eb_relocate_vma_slow(eb, vma);
+ if (ret)
+ goto err;
+ }
}
/* Leave the user relocations as are, this is the painfully slow path,
@@ -1102,21 +1414,67 @@ eb_relocate_slow(struct i915_execbuffer *eb)
*/
err:
- drm_free_large(reloc);
- drm_free_large(reloc_offset);
- return ret;
+ if (ret == -EAGAIN)
+ goto repeat;
+
+out:
+ if (have_copy) {
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
+
+ for (i = 0; i < count; i++) {
+ const struct drm_i915_gem_exec_object2 *entry =
+ &eb->exec[i];
+ struct drm_i915_gem_relocation_entry *relocs;
+
+ if (entry->relocation_count == 0)
+ continue;
+
+ relocs = to_ptr(typeof(*relocs), entry->relocs_ptr);
+ drm_free_large(relocs);
+ }
+ }
+
+ return ret ?: have_copy;
+}
+
+static void eb_export_fence(struct drm_i915_gem_object *obj,
+ struct drm_i915_gem_request *req,
+ unsigned int flags)
+{
+ struct reservation_object *resv = obj->resv;
+
+ /* Ignore errors from failing to allocate the new fence, we can't
+ * handle an error right now. Worst case should be missed
+ * synchronisation leading to rendering corruption.
+ */
+ reservation_object_lock(resv, NULL);
+ if (flags & EXEC_OBJECT_WRITE)
+ reservation_object_add_excl_fence(resv, &req->fence);
+ else if (reservation_object_reserve_shared(resv) == 0)
+ reservation_object_add_shared_fence(resv, &req->fence);
+ reservation_object_unlock(resv);
+}
+
+static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
+{
+ return !(obj->cache_level == I915_CACHE_NONE ||
+ obj->cache_level == I915_CACHE_WT);
}
static int
eb_move_to_gpu(struct i915_execbuffer *eb)
{
- struct i915_vma *vma;
+ const unsigned int count = eb->buffer_count;
+ unsigned int i;
int ret;
- list_for_each_entry(vma, &eb->vmas, exec_link) {
+ for (i = 0; i < count; i++) {
+ const struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+ struct i915_vma *vma = to_ptr(struct i915_vma, entry->rsvd2);
struct drm_i915_gem_object *obj = vma->obj;
- if (vma->exec_entry->flags & EXEC_OBJECT_CAPTURE) {
+ if (entry->flags & EXEC_OBJECT_CAPTURE) {
struct i915_gem_capture_list *capture;
capture = kmalloc(sizeof(*capture), GFP_KERNEL);
@@ -1128,8 +1486,8 @@ eb_move_to_gpu(struct i915_execbuffer *eb)
eb->request->capture_list = capture;
}
- if (vma->exec_entry->flags & EXEC_OBJECT_ASYNC)
- continue;
+ if (entry->flags & EXEC_OBJECT_ASYNC)
+ goto skip_flushes;
if (obj->base.write_domain & I915_GEM_DOMAIN_CPU) {
i915_gem_clflush_object(obj, 0);
@@ -1137,10 +1495,33 @@ eb_move_to_gpu(struct i915_execbuffer *eb)
}
ret = i915_gem_request_await_object
- (eb->request, obj, vma->exec_entry->flags & EXEC_OBJECT_WRITE);
+ (eb->request, obj, entry->flags & EXEC_OBJECT_WRITE);
if (ret)
return ret;
+
+skip_flushes:
+ obj->base.write_domain = 0;
+ if (entry->flags & EXEC_OBJECT_WRITE) {
+ obj->base.read_domains = 0;
+ if (!obj->cache_dirty && gpu_write_needs_clflush(obj))
+ obj->cache_dirty = true;
+ intel_fb_obj_invalidate(obj, ORIGIN_CS);
+ }
+ obj->base.read_domains |= I915_GEM_GPU_DOMAINS;
+
+ i915_vma_move_to_active(vma, eb->request, entry->flags);
+ __eb_unreserve_vma(vma, entry);
+ vma->exec_entry = NULL;
+ }
+
+ for (i = 0; i < count; i++) {
+ const struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+ struct i915_vma *vma = to_ptr(struct i915_vma, entry->rsvd2);
+
+ eb_export_fence(vma->obj, eb->request, entry->flags);
+ i915_vma_put(vma);
}
+ eb->exec = NULL;
/* Unconditionally flush any chipset caches (for streaming writes). */
i915_gem_chipset_flush(eb->i915);
@@ -1172,114 +1553,10 @@ i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
return true;
}
-static int
-validate_exec_list(struct drm_device *dev,
- struct drm_i915_gem_exec_object2 *exec,
- int count)
-{
- unsigned relocs_total = 0;
- unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
- unsigned invalid_flags;
- int i;
-
- /* INTERNAL flags must not overlap with external ones */
- BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
-
- invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
- if (USES_FULL_PPGTT(dev))
- invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
-
- for (i = 0; i < count; i++) {
- char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
- int length; /* limited by fault_in_pages_readable() */
-
- if (exec[i].flags & invalid_flags)
- return -EINVAL;
-
- /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
- * any non-page-aligned or non-canonical addresses.
- */
- if (exec[i].flags & EXEC_OBJECT_PINNED) {
- if (exec[i].offset !=
- gen8_canonical_addr(exec[i].offset & PAGE_MASK))
- return -EINVAL;
- }
-
- /* From drm_mm perspective address space is continuous,
- * so from this point we're always using non-canonical
- * form internally.
- */
- exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
-
- if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
- return -EINVAL;
-
- /* pad_to_size was once a reserved field, so sanitize it */
- if (exec[i].flags & EXEC_OBJECT_PAD_TO_SIZE) {
- if (offset_in_page(exec[i].pad_to_size))
- return -EINVAL;
- } else {
- exec[i].pad_to_size = 0;
- }
-
- /* First check for malicious input causing overflow in
- * the worst case where we need to allocate the entire
- * relocation tree as a single array.
- */
- if (exec[i].relocation_count > relocs_max - relocs_total)
- return -EINVAL;
- relocs_total += exec[i].relocation_count;
-
- length = exec[i].relocation_count *
- sizeof(struct drm_i915_gem_relocation_entry);
- /*
- * We must check that the entire relocation array is safe
- * to read, but since we may need to update the presumed
- * offsets during execution, check for full write access.
- */
- if (!access_ok(VERIFY_WRITE, ptr, length))
- return -EFAULT;
-
- if (likely(!i915.prefault_disable)) {
- if (fault_in_pages_readable(ptr, length))
- return -EFAULT;
- }
- }
-
- return 0;
-}
-
-static int eb_select_context(struct i915_execbuffer *eb)
-{
- unsigned int ctx_id = i915_execbuffer2_get_context_id(*eb->args);
- struct i915_gem_context *ctx;
-
- ctx = i915_gem_context_lookup(eb->file->driver_priv, ctx_id);
- if (unlikely(IS_ERR(ctx)))
- return PTR_ERR(ctx);
-
- if (unlikely(i915_gem_context_is_banned(ctx))) {
- DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
- return -EIO;
- }
-
- eb->ctx = i915_gem_context_get(ctx);
- eb->vm = ctx->ppgtt ? &ctx->ppgtt->base : &eb->i915->ggtt.base;
-
- return 0;
-}
-
-static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
-{
- return !(obj->cache_level == I915_CACHE_NONE ||
- obj->cache_level == I915_CACHE_WT);
-}
-
void i915_vma_move_to_active(struct i915_vma *vma,
struct drm_i915_gem_request *req,
unsigned int flags)
{
- struct drm_i915_gem_object *obj = vma->obj;
const unsigned int idx = req->engine->id;
lockdep_assert_held(&req->i915->drm.struct_mutex);
@@ -1293,17 +1570,17 @@ void i915_vma_move_to_active(struct i915_vma *vma,
* *last*.
*/
if (!i915_vma_is_active(vma))
- obj->active_count++;
+ vma->obj->active_count++;
i915_vma_set_active(vma, idx);
i915_gem_active_set(&vma->last_read[idx], req);
list_move_tail(&vma->vm_link, &vma->vm->active_list);
if (flags & EXEC_OBJECT_WRITE) {
+ struct drm_i915_gem_object *obj = vma->obj;
+
if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
i915_gem_active_set(&obj->frontbuffer_write, req);
- /* update for the implicit flush after a batch */
- obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
if (!obj->cache_dirty && gpu_write_needs_clflush(obj))
obj->cache_dirty = true;
}
@@ -1312,42 +1589,6 @@ void i915_vma_move_to_active(struct i915_vma *vma,
i915_gem_active_set(&vma->last_fence, req);
}
-static void eb_export_fence(struct drm_i915_gem_object *obj,
- struct drm_i915_gem_request *req,
- unsigned int flags)
-{
- struct reservation_object *resv = obj->resv;
-
- /* Ignore errors from failing to allocate the new fence, we can't
- * handle an error right now. Worst case should be missed
- * synchronisation leading to rendering corruption.
- */
- reservation_object_lock(resv, NULL);
- if (flags & EXEC_OBJECT_WRITE)
- reservation_object_add_excl_fence(resv, &req->fence);
- else if (reservation_object_reserve_shared(resv) == 0)
- reservation_object_add_shared_fence(resv, &req->fence);
- reservation_object_unlock(resv);
-}
-
-static void
-eb_move_to_active(struct i915_execbuffer *eb)
-{
- struct i915_vma *vma;
-
- list_for_each_entry(vma, &eb->vmas, exec_link) {
- struct drm_i915_gem_object *obj = vma->obj;
-
- obj->base.write_domain = 0;
- if (vma->exec_entry->flags & EXEC_OBJECT_WRITE)
- obj->base.read_domains = 0;
- obj->base.read_domains |= I915_GEM_GPU_DOMAINS;
-
- i915_vma_move_to_active(vma, eb->request, vma->exec_entry->flags);
- eb_export_fence(obj, eb->request, vma->exec_entry->flags);
- }
-}
-
static int
i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
{
@@ -1359,16 +1600,16 @@ i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
return -EINVAL;
}
- cs = intel_ring_begin(req, 4 * 3);
+ cs = intel_ring_begin(req, 4 * 2 + 2);
if (IS_ERR(cs))
return PTR_ERR(cs);
+ *cs++ = MI_LOAD_REGISTER_IMM(4);
for (i = 0; i < 4; i++) {
- *cs++ = MI_LOAD_REGISTER_IMM(1);
*cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
*cs++ = 0;
}
-
+ *cs++ = MI_NOOP;
intel_ring_advance(req, cs);
return 0;
@@ -1404,10 +1645,11 @@ static struct i915_vma *eb_parse(struct i915_execbuffer *eb, bool is_master)
goto out;
vma->exec_entry =
- memset(&eb->shadow_exec_entry, 0, sizeof(*vma->exec_entry));
+ memset(&eb->exec[eb->buffer_count++],
+ 0, sizeof(*vma->exec_entry));
vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
- i915_gem_object_get(shadow_batch_obj);
- list_add_tail(&vma->exec_link, &eb->vmas);
+ vma->exec_entry->rsvd2 = (uintptr_t)vma;
+ i915_vma_get(vma);
out:
i915_gem_object_unpin_pages(shadow_batch_obj);
@@ -1423,7 +1665,7 @@ add_to_client(struct drm_i915_gem_request *req,
}
static int
-execbuf_submit(struct i915_execbuffer *eb)
+eb_submit(struct i915_execbuffer *eb)
{
int ret;
@@ -1449,8 +1691,6 @@ execbuf_submit(struct i915_execbuffer *eb)
if (ret)
return ret;
- eb_move_to_active(eb);
-
return 0;
}
@@ -1542,20 +1782,21 @@ i915_gem_do_execbuffer(struct drm_device *dev,
int out_fence_fd = -1;
int ret;
- if (!i915_gem_check_execbuffer(args))
- return -EINVAL;
-
- ret = validate_exec_list(dev, exec, args->buffer_count);
- if (ret)
- return ret;
+ BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
eb.i915 = to_i915(dev);
eb.file = file;
eb.args = args;
+ if ((args->flags & I915_EXEC_NO_RELOC) == 0)
+ args->flags |= __EXEC_HAS_RELOC;
eb.exec = exec;
- eb.need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
+ eb.ctx = NULL;
+ eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
+ if (USES_FULL_PPGTT(eb.i915))
+ eb.invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
reloc_cache_init(&eb.reloc_cache, eb.i915);
+ eb.buffer_count = args->buffer_count;
eb.batch_start_offset = args->batch_start_offset;
eb.batch_len = args->batch_len;
@@ -1601,6 +1842,9 @@ i915_gem_do_execbuffer(struct drm_device *dev,
}
}
+ if (eb_create(&eb))
+ return -ENOMEM;
+
/* Take a local wakeref for preparing to dispatch the execbuf as
* we expect to access the hardware fairly frequently in the
* process. Upon first dispatch, we acquire another prolonged
@@ -1608,60 +1852,41 @@ i915_gem_do_execbuffer(struct drm_device *dev,
* 100ms.
*/
intel_runtime_pm_get(eb.i915);
-
ret = i915_mutex_lock_interruptible(dev);
if (ret)
- goto pre_mutex_err;
+ goto err_rpm;
ret = eb_select_context(&eb);
- if (ret) {
- mutex_unlock(&dev->struct_mutex);
- goto pre_mutex_err;
- }
+ if (unlikely(ret))
+ goto err_unlock;
- if (eb_create(&eb)) {
- i915_gem_context_put(eb.ctx);
- mutex_unlock(&dev->struct_mutex);
- ret = -ENOMEM;
- goto pre_mutex_err;
- }
-
- /* Look up object handles */
ret = eb_lookup_vmas(&eb);
- if (ret)
- goto err;
-
- /* take note of the batch buffer before we might reorder the lists */
- eb.batch = eb_get_batch(&eb);
-
- /* Move the objects en-masse into the GTT, evicting if necessary. */
- ret = eb_reserve(&eb);
- if (ret)
- goto err;
-
- /* The objects are in their final locations, apply the relocations. */
- if (eb.need_relocs)
+ if (likely(!ret && args->flags & __EXEC_HAS_RELOC))
ret = eb_relocate(&eb);
- if (ret) {
- if (ret == -EFAULT) {
- ret = eb_relocate_slow(&eb);
- BUG_ON(!mutex_is_locked(&dev->struct_mutex));
- }
- if (ret)
- goto err;
- }
+ if (ret == -EAGAIN || ret == -EFAULT)
+ ret = eb_relocate_slow(&eb);
+ if (ret && args->flags & I915_EXEC_NO_RELOC)
+ /* If the user expects the execobject.offset and
+ * reloc.presumed_offset to be an exact match,
+ * as for using NO_RELOC, then we cannot update
+ * the execobject.offset until we have completed
+ * relocation.
+ */
+ args->flags &= ~__EXEC_HAS_RELOC;
+ if (ret < 0)
+ goto err_vma;
/* Set the pending read domains for the batch buffer to COMMAND */
- if (eb.batch->exec_entry->flags & EXEC_OBJECT_WRITE) {
+ if (unlikely(eb.batch->exec_entry->flags & EXEC_OBJECT_WRITE)) {
DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
ret = -EINVAL;
- goto err;
+ goto err_vma;
}
if (eb.batch_start_offset > eb.batch->size ||
eb.batch_len > eb.batch->size - eb.batch_start_offset) {
DRM_DEBUG("Attempting to use out-of-bounds batch\n");
ret = -EINVAL;
- goto err;
+ goto err_vma;
}
if (eb.engine->needs_cmd_parser && eb.batch_len) {
@@ -1670,7 +1895,7 @@ i915_gem_do_execbuffer(struct drm_device *dev,
vma = eb_parse(&eb, drm_is_current_master(file));
if (IS_ERR(vma)) {
ret = PTR_ERR(vma);
- goto err;
+ goto err_vma;
}
if (vma) {
@@ -1696,7 +1921,6 @@ i915_gem_do_execbuffer(struct drm_device *dev,
* batch" bit. Hence we need to pin secure batches into the global gtt.
* hsw should have this fixed, but bdw mucks it up again. */
if (eb.dispatch_flags & I915_DISPATCH_SECURE) {
- struct drm_i915_gem_object *obj = eb.batch->obj;
struct i915_vma *vma;
/*
@@ -1709,10 +1933,10 @@ i915_gem_do_execbuffer(struct drm_device *dev,
* fitting due to fragmentation.
* So this is actually safe.
*/
- vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, 0);
+ vma = i915_gem_object_ggtt_pin(eb.batch->obj, NULL, 0, 0, 0);
if (IS_ERR(vma)) {
ret = PTR_ERR(vma);
- goto err;
+ goto err_vma;
}
eb.batch = vma;
@@ -1748,7 +1972,7 @@ i915_gem_do_execbuffer(struct drm_device *dev,
eb.request->batch = eb.batch;
trace_i915_gem_request_queue(eb.request, eb.dispatch_flags);
- ret = execbuf_submit(&eb);
+ ret = eb_submit(&eb);
err_request:
__i915_add_request(eb.request, ret == 0);
add_to_client(eb.request, file);
@@ -1765,23 +1989,16 @@ i915_gem_do_execbuffer(struct drm_device *dev,
}
err_batch_unpin:
- /*
- * FIXME: We crucially rely upon the active tracking for the (ppgtt)
- * batch vma for correctness. For less ugly and less fragility this
- * needs to be adjusted to also track the ggtt batch vma properly as
- * active.
- */
if (eb.dispatch_flags & I915_DISPATCH_SECURE)
i915_vma_unpin(eb.batch);
-err:
- /* the request owns the ref now */
- eb_destroy(&eb);
+err_vma:
+ eb_release_vma(&eb);
+ i915_gem_context_put(eb.ctx);
+err_unlock:
mutex_unlock(&dev->struct_mutex);
-
-pre_mutex_err:
- /* intel_gpu_busy should also get a ref, so it will free when the device
- * is really idle. */
+err_rpm:
intel_runtime_pm_put(eb.i915);
+ eb_destroy(&eb);
if (out_fence_fd != -1)
put_unused_fd(out_fence_fd);
err_in_fence:
@@ -1808,9 +2025,27 @@ i915_gem_execbuffer(struct drm_device *dev, void *data,
return -EINVAL;
}
+ exec2.buffers_ptr = args->buffers_ptr;
+ exec2.buffer_count = args->buffer_count;
+ exec2.batch_start_offset = args->batch_start_offset;
+ exec2.batch_len = args->batch_len;
+ exec2.DR1 = args->DR1;
+ exec2.DR4 = args->DR4;
+ exec2.num_cliprects = args->num_cliprects;
+ exec2.cliprects_ptr = args->cliprects_ptr;
+ exec2.flags = I915_EXEC_RENDER;
+ i915_execbuffer2_set_context_id(exec2, 0);
+
+ if (!i915_gem_check_execbuffer(&exec2))
+ return -EINVAL;
+
/* Copy in the exec list from userland */
- exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
- exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
+ exec_list = drm_malloc_gfp(args->buffer_count,
+ sizeof(*exec_list),
+ __GFP_NOWARN | GFP_TEMPORARY);
+ exec2_list = drm_malloc_gfp(args->buffer_count + 1,
+ sizeof(*exec2_list),
+ __GFP_NOWARN | GFP_TEMPORARY);
if (exec_list == NULL || exec2_list == NULL) {
DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
args->buffer_count);
@@ -1841,36 +2076,23 @@ i915_gem_execbuffer(struct drm_device *dev, void *data,
exec2_list[i].flags = 0;
}
- exec2.buffers_ptr = args->buffers_ptr;
- exec2.buffer_count = args->buffer_count;
- exec2.batch_start_offset = args->batch_start_offset;
- exec2.batch_len = args->batch_len;
- exec2.DR1 = args->DR1;
- exec2.DR4 = args->DR4;
- exec2.num_cliprects = args->num_cliprects;
- exec2.cliprects_ptr = args->cliprects_ptr;
- exec2.flags = I915_EXEC_RENDER;
- i915_execbuffer2_set_context_id(exec2, 0);
-
ret = i915_gem_do_execbuffer(dev, file, &exec2, exec2_list);
- if (!ret) {
+ if (exec2.flags & __EXEC_HAS_RELOC) {
struct drm_i915_gem_exec_object __user *user_exec_list =
u64_to_user_ptr(args->buffers_ptr);
/* Copy the new buffer offsets back to the user's exec list. */
for (i = 0; i < args->buffer_count; i++) {
+ if ((exec2_list[i].offset & UPDATE) == 0)
+ continue;
+
exec2_list[i].offset =
- gen8_canonical_addr(exec2_list[i].offset);
- ret = __copy_to_user(&user_exec_list[i].offset,
- &exec2_list[i].offset,
- sizeof(user_exec_list[i].offset));
- if (ret) {
- ret = -EFAULT;
- DRM_DEBUG("failed to copy %d exec entries "
- "back to user (%d)\n",
- args->buffer_count, ret);
+ gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
+ exec2_list[i].offset &= PIN_OFFSET_MASK;
+ if (__copy_to_user(&user_exec_list[i].offset,
+ &exec2_list[i].offset,
+ sizeof(user_exec_list[i].offset)))
break;
- }
}
}
@@ -1884,56 +2106,63 @@ i915_gem_execbuffer2(struct drm_device *dev, void *data,
struct drm_file *file)
{
struct drm_i915_gem_execbuffer2 *args = data;
- struct drm_i915_gem_exec_object2 *exec2_list = NULL;
+ struct drm_i915_gem_exec_object2 *exec2_list;
int ret;
if (args->buffer_count < 1 ||
- args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
+ args->buffer_count >= UINT_MAX / sizeof(*exec2_list) - 1) {
DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
return -EINVAL;
}
- exec2_list = drm_malloc_gfp(args->buffer_count,
+ if (!i915_gem_check_execbuffer(args))
+ return -EINVAL;
+
+ exec2_list = drm_malloc_gfp(args->buffer_count + 1,
sizeof(*exec2_list),
- GFP_TEMPORARY);
+ __GFP_NOWARN | GFP_TEMPORARY);
if (exec2_list == NULL) {
DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
args->buffer_count);
return -ENOMEM;
}
- ret = copy_from_user(exec2_list,
- u64_to_user_ptr(args->buffers_ptr),
- sizeof(*exec2_list) * args->buffer_count);
- if (ret != 0) {
- DRM_DEBUG("copy %d exec entries failed %d\n",
- args->buffer_count, ret);
+ if (copy_from_user(exec2_list,
+ u64_to_user_ptr(args->buffers_ptr),
+ sizeof(*exec2_list) * args->buffer_count)) {
+ DRM_DEBUG("copy %d exec entries failed\n", args->buffer_count);
drm_free_large(exec2_list);
return -EFAULT;
}
ret = i915_gem_do_execbuffer(dev, file, args, exec2_list);
- if (!ret) {
- /* Copy the new buffer offsets back to the user's exec list. */
+
+ /* Now that we have begun execution of the batchbuffer, we ignore
+ * any new error after this point. Also given that we have already
+ * updated the associated relocations, we try to write out the current
+ * object locations irrespective of any error.
+ */
+ if (args->flags & __EXEC_HAS_RELOC) {
struct drm_i915_gem_exec_object2 __user *user_exec_list =
- u64_to_user_ptr(args->buffers_ptr);
+ u64_to_user_ptr(args->buffers_ptr);
int i;
+ /* Copy the new buffer offsets back to the user's exec list. */
+ user_access_begin();
for (i = 0; i < args->buffer_count; i++) {
+ if ((exec2_list[i].offset & UPDATE) == 0)
+ continue;
+
exec2_list[i].offset =
- gen8_canonical_addr(exec2_list[i].offset);
- ret = __copy_to_user(&user_exec_list[i].offset,
- &exec2_list[i].offset,
- sizeof(user_exec_list[i].offset));
- if (ret) {
- ret = -EFAULT;
- DRM_DEBUG("failed to copy %d exec entries "
- "back to user\n",
- args->buffer_count);
- break;
- }
+ gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
+ unsafe_put_user(exec2_list[i].offset,
+ &user_exec_list[i].offset,
+ end_user);
}
+end_user:
+ user_access_end();
}
+ args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS;
drm_free_large(exec2_list);
return ret;
}
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 507463f144c9..5097dcc3987a 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -463,7 +463,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
size, obj->base.size,
flags & PIN_MAPPABLE ? "mappable" : "total",
end);
- return -E2BIG;
+ return -ENOSPC;
}
ret = i915_gem_object_pin_pages(obj);
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h
index 88543fafcffc..062addfee6ef 100644
--- a/drivers/gpu/drm/i915/i915_vma.h
+++ b/drivers/gpu/drm/i915/i915_vma.h
@@ -103,6 +103,7 @@ struct i915_vma {
/** This vma's place in the execbuf reservation list */
struct list_head exec_link;
+ struct list_head reloc_link;
/** This vma's place in the eviction list */
struct list_head evict_link;
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
index 14e9c2fbc4e6..5ea373221f49 100644
--- a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
@@ -304,7 +304,7 @@ static int igt_evict_vm(void *arg)
goto cleanup;
/* Everything is pinned, nothing should happen */
- err = i915_gem_evict_vm(&ggtt->base, false);
+ err = i915_gem_evict_vm(&ggtt->base);
if (err) {
pr_err("i915_gem_evict_vm on a full GGTT returned err=%d]\n",
err);
@@ -313,7 +313,7 @@ static int igt_evict_vm(void *arg)
unpin_ggtt(i915);
- err = i915_gem_evict_vm(&ggtt->base, false);
+ err = i915_gem_evict_vm(&ggtt->base);
if (err) {
pr_err("i915_gem_evict_vm on a full GGTT returned err=%d]\n",
err);
--
2.11.0
More information about the Intel-gfx
mailing list